Unmanned aerial vehicle photovoltaic real-time inspection system and method based on end-edge-cloud architecture

Abstract

The invention provides an unmanned aerial vehicle photovoltaic real-time inspection system and method based on an end-edge-cloud architecture, and belongs to the field of photovoltaic power station operation and maintenance. Comprising an end device which directly interacts with an environment or a user and collects data and comprises an unmanned aerial vehicle end, a flight controller, a pan-tilt camera end and a mobile phone end; the edge computing platform is realized by adopting an embedded platform carried above the unmanned aerial vehicle end, and is used for reading end equipment data, executing fuzzy judgment based on a visible light image and two-stage defect detection based on an infrared image, and uploading an image detection result to the cloud server; and the cloud server is responsible for data storage and message pushing, receiving the image result uploaded by the embedded platform and pushing the image result to the mobile phone terminal. According to the invention, an end-edge-cloud collaborative architecture facing a power station is provided, the diagnosis and decision delay is greatly shortened, the rapid identification and positioning of the defects of the photovoltaic equipment are realized, and the high-precision identification of the defects is realized by adopting a two-stage detection algorithm based on target identification and defect classification.

Description

Technical Field

[0001] The present invention belongs to the field of photovoltaic power station operation and maintenance, and in particular relates to a drone photovoltaic real-time inspection system and method based on an end-edge-cloud architecture. Background Art

[0002] As the global energy crisis intensifies and severe climate change accelerates, energy transition is crucial. Photovoltaic power generation, as a clean, renewable energy source, is rapidly developing globally. The scale of photovoltaic power station construction in western my country continues to expand, gradually moving toward higher altitudes and larger scales. The commissioning of projects such as the Xingchuan Photovoltaic Power Station in Ganzi, Sichuan, is of great significance to the local clean energy supply and economic development. However, factors such as low oxygen levels, large temperature fluctuations, and unpredictable weather in plateau environments can easily cause failures in photovoltaic modules and inverters, significantly increasing the difficulty of operation and maintenance. Achieving efficient operation and maintenance and fault diagnosis of photovoltaic systems in high-altitude environments with complex terrain has become a major challenge that the photovoltaic industry urgently needs to address.

[0003] Traditional manual inspections present numerous problems in high-altitude areas, including heavy workloads and low efficiency. Furthermore, the lack of oxygen in the environment not only reduces inspection efficiency but can also cause permanent damage to the physical and mental health of maintenance personnel, posing a safety hazard. In recent years, drone inspections have garnered widespread attention in the field of operation and maintenance due to their high flexibility, maneuverability, and programmability. In the case of drone photovoltaic inspections, infrared image defect recognition often requires offline processing at ground stations, leading to lengthy data analysis cycles and difficulty meeting the need for real-time defect response. Furthermore, drones require high image accuracy when detecting photovoltaic infrared image defects. Using video stream detection can easily lead to misjudgment of defects due to image blur. Existing real-time video stream detection systems are not suitable for drone inspections.

[0004] Real-time inspection systems require lightweight models for efficient detection while ensuring accuracy. Users also need timely access to test results. Therefore, a new technology is needed to address the operational challenges of photovoltaic power plants in high-altitude areas, enabling efficient maintenance and fault diagnosis, meeting real-time response requirements, and reducing costs and technical barriers. Summary of the Invention

[0005] In order to solve the above problems, the present invention proposes a drone photovoltaic real-time inspection system and method based on the end-edge-cloud architecture, and proposes an end-edge-cloud collaborative architecture for power stations, which greatly shortens the diagnosis and decision-making delays, realizes the rapid identification and positioning of photovoltaic equipment defects, and adopts a two-stage detection algorithm based on target recognition and defect classification to achieve high-precision defect identification.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] In the first aspect, the present invention proposes a real-time photovoltaic inspection system using a drone based on an edge-cloud architecture, comprising:

[0008] The terminal device directly interacts with the environment or the user and collects data, including the drone terminal, flight controller, gimbal camera terminal and mobile phone terminal; the drone terminal is equipped with a thermal imaging camera, which collects infrared and visible light images of the photovoltaic strings under the control of the flight controller, and measures and maps the working area. The collected data is stored in the gimbal camera terminal;

[0009] The edge computing platform is implemented using an embedded platform mounted on the drone. It is used to read data from the gimbal camera, perform fuzzy judgment based on visible light images and two-stage defect detection based on infrared images, and upload image detection results to the cloud server.

[0010] The cloud server is responsible for data storage and message push, receiving image results uploaded by the embedded platform and pushing them to the mobile phone.

[0011] As a preferred embodiment of the present invention, the embedded platform adopts a file paging download strategy for multiple image metadata to read the PTZ camera end data, specifically:

[0012] Set input: file list, JSON file path, maximum file size and latest date;

[0013] Set output: add target file;

[0014] The specific process is as follows: initialize the JSON structure for storing historical file records and load the existing JSON records from the JSON file path; through multiple paging queries, traverse each file name, file timestamp and file size in the file list. If the file timestamp is earlier than the most recent date, or the file size exceeds the maximum file size, or the file name suffix is not the suffix "_T" for infrared images or the suffix "_Z" for visible light images, it will not be read; if the read file is not recorded in the existing JSON file, the file metadata is added to the JSON record and also added to the newly added target file; finally, save the updated JSON record to the JSON file path and output the newly added target file.

[0015] As a preferred embodiment of the present invention, the embedded platform has built-in image blur detection algorithm and two-stage defect detection algorithm.

[0016] As a preferred embodiment of the present invention, the image blur detection algorithm is specifically:

[0017] Set input: image path and blur threshold;

[0018] Set output: image blur state and processed image;

[0019] The specific process is as follows: first, read the visible light image from the image path and convert it into a grayscale image; then calculate the Laplacian variance of the grayscale image as the fuzzy score; if the fuzzy score is less than the fuzzy threshold, set the fuzzy state of the corresponding infrared image to TRUE, generate JSON annotation information to record the fuzzy image, and no longer perform defect detection; if the fuzzy score is greater than the fuzzy threshold, set the fuzzy state to FALSE and call the two-stage defect detection algorithm.

[0020] As a preferred embodiment of the present invention, the two-stage defect detection algorithm is specifically as follows:

[0021] Taking infrared images as input, we first use a target recognition network to identify each photovoltaic module in the infrared image, and crop each sub-image containing the module based on the recognition results; then a classification network is used to identify the module defect types, which include normal modules, hot spot modules, severe hot spot modules, diode conduction modules, abnormal heating modules, distorted modules, and power-limited modules.

[0022] As a preferred embodiment of the present invention, the target recognition network adopts the YOLOv11s network, and the classification network adopts the ResNet34 network.

[0023] As a preferred embodiment of the present invention, the two-stage defect detection algorithm is trained in stages. In the first stage, a target recognition network is trained, with an infrared image as input and an annotated photovoltaic module block diagram as a label, to identify the position of each photovoltaic module in the infrared image; in the second stage, a classification network is trained to identify the defect type of the photovoltaic module sub-image with a separate photovoltaic module sub-image as input and an annotated photovoltaic module defect type as a label.

[0024] As a preferred embodiment of the present invention, when the embedded platform uploads the image detection result to the cloud server, if the image size is greater than a threshold, a compression operation is performed on the image.

[0025] As a preferred embodiment of the present invention, when the cloud server sends the result to the mobile phone, it constructs a JSON message containing the image address and annotation information; uses MQTT to publish the JSON message to a specific topic; the mobile phone subscribes to the topic and receives the message, parses the received JSON message, extracts the image and annotation information of the image, downloads the image file corresponding to the address through an HTTP request, and displays the image and annotation data.

[0026] The beneficial effects of the present invention are:

[0027] The present invention is based on end-edge-cloud collaborative photovoltaic real-time inspection technology and a two-stage detection method. It is superior to the traditional offline inspection mode in detection accuracy and speed, reduces dependence on traditional ground station servers, and effectively reduces the operation and maintenance costs of power stations. It provides reliable technical support for real-time inspection and intelligent operation and maintenance of high-altitude photovoltaic power stations under complex climatic conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a framework diagram of the UAV photovoltaic real-time inspection system based on the end-edge-cloud architecture proposed by the present invention;

[0029] Figure 2 is the blur score result of the clear, blur-free picture in this embodiment;

[0030] Figure 3 is the surveying and mapping result of this embodiment;

[0031] Figure 4 This is the inspection machine position diagram of this embodiment;

[0032] Figure 5 is the marking result of the frame where the photovoltaic module is located in this embodiment;

[0033] Figure 6 Schematic diagram of defect types of photovoltaic modules in this embodiment. DETAILED DESCRIPTION

[0034] In order to make the above-mentioned objects, features and advantages of the present invention more clearly understood, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. The technical features in the various embodiments of the present invention can be combined accordingly without conflicting with each other.

[0035] like Figure 1 As shown, the present invention designs and implements an intelligent operation and maintenance system for real-time drone inspection of photovoltaic power stations based on an end-edge-cloud architecture to replace traditional manual inspections and high-cost manually controlled drone inspections.

[0036] (1) Intelligent operation and maintenance system framework based on end-edge-cloud collaborative architecture

[0037] This invention addresses the complex climate, redundant data, and high-efficiency requirements faced by photovoltaic operation and maintenance in high-altitude areas. It proposes an end-edge-cloud collaborative architecture for power plants. End devices refer to devices that directly interact with the environment or users and collect data; edge computing platforms refer to devices or platforms with sufficient computing power located close to the data source; and cloud servers are virtualized servers based on cloud computing technology. This architecture, serving as the foundation for real-time drone photovoltaic inspections, can significantly reduce diagnostic and decision-making latency, enabling rapid identification and location of photovoltaic equipment defects.

[0038] Specifically, the end device, edge computing platform, and cloud server based on the end-edge-cloud architecture are composed as follows:

[0039] (1) End device

[0040] The system includes the drone, flight controller, gimbal camera, and mobile phone. It is responsible for collecting photovoltaic panel image data. For example, a drone equipped with a thermal imaging camera, under the control of the flight controller, quickly and extensively collects infrared and visible light images of the photovoltaic panels, as well as measures and maps the work area. The collected data is stored on the gimbal camera. The flight controller can send control commands to the drone and plan the drone's path, enabling human-machine interaction between the flight controller and the mobile phone. The gimbal camera transmits the data to the edge computing platform.

[0041] In this example, a DJI drone model is used. It has a flight time of approximately one hour, high positioning accuracy, can be equipped with a high-definition visible light camera and thermal imaging lens, and has multiple payload interfaces. The H30T thermal infrared camera is mounted below the drone. It uses a high-resolution camera and captures visible light and infrared images that are stored on the gimbal camera.

[0042] (2) Edge computing platform

[0043] An embedded platform mounted on top of the drone is used, which has a built-in lightweight small-model neural network. It reads and downloads real-time images taken by the drone, performs fuzzy judgment and defect detection, and uploads the results to the cloud server.

[0044] In this embodiment, the embedded platform is mounted above the drone, and the exterior uses a 3D-printed shell that is adapted to the drone structure. The interior contains multi-functional modules, such as an artificial intelligence developer kit, an E-Port adapter module, a DC-DC power supply module, a WiFi communication module and a fixed bracket, and is equipped with a hardware protection box for severe weather.

[0045] The embedded platform designs file download management and image blur detection algorithms in a multilingual collaborative manner. Here, the original images collected by a drone equipped with an H30T thermal infrared camera include infrared images and visible light images. In embedded development platforms, real-time data download processing requires a large amount of bandwidth and storage resources, so developing an efficient file management strategy is particularly important. This paper designs a file paging download strategy that combines multiple image metadata. By dynamically recording and updating the file list, only newly added files are downloaded, effectively optimizing the file download process and reducing resource waste.

[0046] In a specific implementation of the present invention, the file paging download strategy combining multiple image metadata is specifically as follows:

[0047] Set input: file list, JSON file path, maximum file size and latest date;

[0048] Set output: Add new target file.

[0049] First, initialize the JSON structure for storing historical file records, and then load the existing JSON records from the JSON file path. Subsequently, through multiple paging queries, traverse each file name, file timestamp, and file size in the file list. If the file timestamp is earlier than the most recent date, or the file size exceeds the maximum file size, or the file name suffix is not "_T" or "_Z", it will not be read. When a file is not recorded in an existing JSON file, the file's metadata will be added to the JSON record and also to the newly added target file. Finally, save the updated JSON record to the JSON file path and output the newly added target file.

[0050] The present invention adopts the above-mentioned paging download strategy combining timestamp, file size and file name. First, the paging range and circular query are set to obtain the file list in batches from the gimbal camera end, reduce single communication requests, and improve the stability of the embedded system. Secondly, the timestamp of the file is used to determine whether it belongs to the range of the last seven days, so as to eliminate old files. Large files are further filtered by limiting the file size to prevent batch files from being downloaded by mistake. Finally, the file name suffix is combined to ensure that only infrared images ("_T") and visible light images ("_Z") taken by the drone are downloaded. In addition, the above-mentioned dynamic update mechanism of the file list based on the JSON structure is adopted to avoid repeated downloading of existing files by storing historical file information, thereby realizing the gradual incremental update function.

[0051] In a specific implementation of the present invention, an image blur detection algorithm is designed to achieve blur judgment.

[0052] In real-time drone photovoltaic inspections, image blur detection algorithms are crucial for ensuring accurate inspection results. Aerial drone images captured in complex environments are susceptible to distortion due to motion blur or strong winds. Large areas of undetected strings can be clearly seen in the lower left corner of severely blurred images, severely impacting subsequent defect detection results. To address this issue, after downloading the image file, we designed an image blur detection algorithm based on the Laplacian operator, as follows:

[0053] Set input: image path and blur threshold;

[0054] Set the output: the image blur state and the processed image.

[0055] The specific process is as follows: First, the image is read from the image path and the visible light image is converted to a grayscale image. Then, the Laplacian variance of the grayscale image is calculated as the fuzzy score. If the fuzzy score is less than the fuzzy threshold, the fuzzy state is set to TRUE, a JSON annotation is generated to record the blurred image, and defect detection is no longer performed. If the fuzzy score is greater than or equal to the fuzzy threshold, the fuzzy state is set to FALSE, the defect detection algorithm is invoked, and a processed image with the detection results is returned. After image compression, it is uploaded to the cloud server.

[0056] To reduce the amount of calculation, the present invention evaluates clarity by calculating the variance of the grayscale image corresponding to the visible light image. When the blur score is lower than the set threshold, the image is deemed unusable, the detection stage is skipped, and the image is directly uploaded to the cloud and marked with the "image blur" label. In terms of the selection of the blur threshold, this embodiment calculates the blur score of 316 images taken on-site at a high-altitude photovoltaic power station and manually judged to be clear and free of blur. Figure 2 As shown in the figure, when the blur threshold is higher than 800, the image is considered normal and defect detection can be performed. This algorithm effectively ensures data quality and improves the accuracy and reliability of subsequent image detection.

[0057] In a specific implementation of the present invention, an image uplink algorithm is designed to upload relevant data from the embedded platform to the cloud server and complete the storage operation, specifically:

[0058] Set input: file path and annotation data;

[0059] Set output: HTTP response status code.

[0060] The specific process is as follows: First, check the image size corresponding to the file path. If the image is larger than the preset size (for example, 1MB), perform image compression. Package the file path and annotation data using the multipart / form-data format. After packaging, upload these data to the pre-defined interface of the cloud server (for example, port 8082) through an HTTP POST request, receive the code returned by the cloud, and record the upload information. After receiving the request, the cloud server will parse the image file and annotation information from it, and call the InsertImage() function to save the image to the specified folder; then call the InsertImageLable() function to overwrite the annotation data into the annotation table. After completing these operations, the cloud server returns an HTTP response status code to indicate that the data has been successfully stored. Data that does not exceed the preset size does not need to be compressed and can be uploaded directly.

[0061] Once the embedded platform completes image processing, it can immediately upload the results to the cloud for data storage. This step allows the cloud server to have a clear record of the results before they are downloaded and subsequently processed.

[0062] (3) Cloud Server

[0063] Responsible for data storage and message push, receiving photos uploaded by the embedded platform and pushing them to the mobile phone, displaying the test results.

[0064] After data is stored in the cloud server, the cloud uses an MQTT messaging server to send update notifications to mobile devices that have subscribed to the corresponding topic. The notifications provide the image access path and annotation information in the form of a URL. The mobile device then requests the image from the cloud via HTTP and displays it.

[0065] In a specific implementation of the present invention, an image downlink algorithm is designed to achieve real-time notification and data downlink from the cloud server to the mobile device, specifically:

[0066] Set input: image address, annotation information;

[0067] Set output: The image detection result pops up on the mobile phone.

[0068] The specific process is as follows: construct a JSON message containing the image address and annotation information; use MQTT to publish the JSON message to a specific topic; the mobile phone subscribes to the topic and receives the message, parses the received JSON message, extracts the image and annotation information from it, and can further download the image file corresponding to the address through HTTP request to display the image and annotation data.

[0069] Leveraging MQTT notifications and HTTP download mechanisms, cloud-based annotations and images are efficiently delivered to mobile devices. Leveraging the cloud's data fusion and messaging capabilities, this significantly shortens the fault diagnosis and response cycle. This eliminates the need for mobile devices to poll for data updates or receive large image files in messages. The cloud can issue notifications quickly and efficiently during data transmission. This process not only reduces the network burden on mobile devices but also improves resource utilization and overall system responsiveness.

[0070] (2) Communication link of intelligent operation and maintenance system framework

[0071] When the drone performs routine inspection tasks, during its first flight, the operation and maintenance personnel need to control the remote control to plan the drone's path and send shooting commands to complete the mapping task. The drone flies according to the planned track points, and the thermal infrared lens completes the task of collecting infrared and visible light images, achieving efficient and high-quality acquisition of original images. The embedded platform designs file download management and image blur detection algorithms in a multi-language collaborative manner, and deploys a defect detection model, which traverses the downloaded images and performs blur judgment and defect detection. Among them, images that fail the blur judgment are directly uploaded to the cloud server with blur marks, and no defect detection is performed. After the mobile phone sees the blur mark, it can resend the shooting command. Further defect detection is performed on the images that pass the blur judgment, completing the detection and reasoning of photovoltaic infrared defects, and uploading the results to the cloud server. Using the cloud-based messaging function, the cloud server pushes the results to the mobile phone through the messaging server. The operation and maintenance personnel can view them in real time and obtain the inspection report immediately after the flight, which greatly shortens the response cycle of defect diagnosis.

[0072] The present invention builds a complete communication link, realizes efficient file management through paged download and JSON dynamic update, adopts Laplacian operator for image blur detection to ensure image quality, and completes the whole process from data acquisition and screening to blur detection, deep learning reasoning, and result transmission and display.

[0073] It should be noted that the drone's flight and shooting missions do not require real-time manual control. During the first flight, the drone needs to be controlled to patrol and map the target area, plan the inspection position and route, and then automatically perform the inspection mission according to the plan.

[0074] This example uses the Xingchuan PV power station in Sichuan as an example. 21 subarrays were inspected. The subarrays are located in a flat, single-axis area. Each subarray generates an average of 3 MW, for a total inspection capacity of 63 MW. A training dataset was constructed using data collected from drone inspections. A two-stage defect detection method was used to identify PV module defects. The time and cost of different inspection methods were also compared.

[0075] S1, flight mapping

[0076] A drone equipped with an H30T thermal imaging camera was used to perform mapping tasks on 21 sub-arrays. Manual path planning was performed to ensure that the drone could fully cover each sub-array. The mapping results are as follows: Figure 3 shown.

[0077] S2, image acquisition, construction of training set

[0078] Based on the surveying and mapping, inspection images were collected for 21 sub-arrays, and a total of 949 camera positions were recorded, such as Figure 4 As shown in the figure, the flight altitude is set to 60 meters to ensure that the collected infrared data set is clear and complete and can accurately reflect the status of the photovoltaic modules.

[0079] In this embodiment, 612 training images of 15 sub-arrays were selected from 949 infrared data sets for training set construction. When constructing the training set, the training set images need to be annotated. The annotation content is the box where each photovoltaic component in the photovoltaic string in the infrared image is located. The annotation results are as follows: Figure 5 As shown in the figure, each PV panel corresponds to a sub-image in the infrared image. Due to the long-term power outages at the power plant, the power generation within the strings is uneven, resulting in uneven brightness in the infrared image. Therefore, special attention should be paid to the accuracy of the annotations.

[0080] Each labeled PV module sub-image is classified into normal modules, hot spot modules, severe hot spot modules, diode conduction modules, abnormal heating modules, twisted modules, and power-limited (mildly twisted) modules, and the number of modules of each type is counted, as shown in Table 1.

[0081] Table 1

[0082] category Quantity (sheets) Normal components 211 Hot spot components 121 Severe hot spot components 25 Diode conduction components 124 Abnormal components 122 Distortion Component 168 Power-limited (mildly distorted) components 125

[0083] Among them, the number of severe hot spot components is relatively small, and data enhancement technology is required for image augmentation. The augmentation methods include left-right flipping (probability of occurrence 0.9), vertical rotation (probability of occurrence 0.2), brightness conversion (probability of occurrence 0.1), and image blurring (probability of occurrence 0.1), which will increase the 25 samples of severe hot spot components to 125. Figure 6 Shows examples of subgraphs for each type of component.

[0084] S3 trains a two-stage defect detection model, which first detects components and then identifies defects. This model includes a YOLOv11s network for component identification and a ResNet network for component classification.

[0085] By comparing grayscale images of infrared and visible light images for detection, this paper found that infrared image input offers better and more stable detection performance than grayscale image data under various IOU requirements. This is because, while grayscale images can mitigate the effects of uneven brightness within photovoltaic modules, they only retain information in a single intensity channel, weakening its perception and having the opposite effect. Therefore, this paper uses the original infrared image as input.

[0086] In this embodiment, the YOLOv11s model is first used to detect the components. The infrared image is used as input, and the YOLOv11s model is trained to identify each photovoltaic component in the infrared image.

[0087] Based on the detection results of the YOLOv11s network, a sub-image of each photovoltaic module was cropped from the infrared image, and then a ResNet34 network was used to classify component defects. Here, each photovoltaic module sub-image was used as input, and the ResNet34 network was trained to identify the sub-image type of each photovoltaic module.

[0088] In this embodiment, the trained YOLOv11s network has an accuracy rate of 99.5% in identifying photovoltaic modules, and the ResNet34 network has an accuracy rate of 94.8% in classifying photovoltaic module types.

[0089] In step S4, the drone performs routine inspection tasks, flies according to the planned track points, completes the task of collecting infrared and visible light images at the target position, and stores them on the gimbal camera side. The embedded platform reads the data from the gimbal camera side, first converts the visible light image into a grayscale image, and calculates the Laplacian variance of the grayscale image as a fuzzy score. If the fuzzy score is less than the fuzzy threshold, the corresponding infrared image is marked as fuzzy and uploaded directly to the cloud server without performing defect detection. Otherwise, the two-stage defect detection algorithm is called to identify the defect type of each photovoltaic module in the infrared image. The image detection results are uploaded to the cloud server. The cloud server pushes the results to the mobile phone side through the message server, and an inspection report is generated after the flight mission is completed.

[0090] In this experiment, drone inspection time was measured, with an average single-image inspection time of 5.74 seconds. This time was primarily affected by the time required to read and download images from the drone. Specifically, after the image stream is captured by the H30T, it is first stored on the SD card. Only after the drone reads the file can the embedded platform detect the updated file list. This two-step hardware transfer process takes a considerable amount of time, but it still ensures that the drone can complete the inspection task within the interval between any two photovoltaic image data acquisitions.

[0091] This two-stage defect detection method, based on YOLOv11s and ResNet, achieves high-precision identification of photovoltaic module defects. Compared with traditional offline drone inspection solutions and manual inspections, real-time drone inspection technology not only significantly shortens inspection and detection time but also effectively reduces operation and maintenance costs through economical embedded configuration. Furthermore, this technology enables operators to monitor the progress of drone footage and module fault status in real time, significantly improving the response speed and decision-making quality of power plant operations and maintenance.

[0092] It should also be noted that those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working process of the system described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here. In the various embodiments provided in this application, the division of steps or modules in the system and method is only a logical function division. In actual implementation, there may be other division methods, for example, multiple modules or steps can be combined or integrated together, and a module or step can also be split.

[0093] The embodiment described above is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Persons skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, any technical solution obtained by equivalent substitution or equivalent transformation falls within the scope of protection of the present invention.

Claims

1. A UAV photovoltaic real-time inspection system based on the end-edge-cloud architecture, characterized by: include: The terminal device directly interacts with the environment or the user and collects data, including the drone terminal, flight controller, gimbal camera terminal and mobile phone terminal; the drone terminal is equipped with a thermal imaging camera, which collects infrared and visible light images of the photovoltaic strings under the control of the flight controller, and measures and maps the working area. The collected data is stored in the gimbal camera terminal; The edge computing platform is implemented using an embedded platform mounted on the drone. It is used to read data from the gimbal camera, perform fuzzy judgment based on visible light images and two-stage defect detection based on infrared images, and upload image detection results to the cloud server. The cloud server is responsible for data storage and message push, receiving image results uploaded by the embedded platform and pushing them to the mobile phone.

2. The UAV photovoltaic real-time inspection system based on the end-edge-cloud architecture according to claim 1 is characterized in that: The embedded platform uses a file paging download strategy for multiple image metadata to read the data from the gimbal camera. Specifically: Set input: file list, JSON file path, maximum file size and latest date; Set output: add target file; The specific process is as follows: initialize the JSON structure to store historical file records, and load the existing JSON records from the JSON file path; Through multiple paged queries, each file name, file timestamp, and file size in the file list is traversed. If the file timestamp is earlier than the most recent date, or the file size exceeds the maximum file size, or the file name suffix is not "_T" for infrared images or "_Z" for visible light images, the file will not be read. If the read file is not recorded in the existing JSON file, the file metadata is added to the JSON record and also to the newly added target file. Finally, the updated JSON record is saved to the JSON file path and the newly added target file is output.

3. The UAV photovoltaic real-time inspection system based on the end-edge-cloud architecture according to claim 1 is characterized in that: The embedded platform has built-in image blur detection algorithm and two-stage defect detection algorithm.

4. The UAV photovoltaic real-time inspection system based on the end-edge-cloud architecture according to claim 3 is characterized in that: The image blur detection algorithm is specifically as follows: Set input: image path and blur threshold; Set output: image blur state and processed image; The specific process is as follows: first, read the visible light image from the image path and convert it into a grayscale image; then calculate the Laplacian variance of the grayscale image as the fuzzy score; if the fuzzy score is less than the fuzzy threshold, set the fuzzy state of the corresponding infrared image to TRUE, generate JSON annotation information to record the fuzzy image, and no longer perform defect detection; if the fuzzy score is greater than the fuzzy threshold, set the fuzzy state to FALSE and call the two-stage defect detection algorithm.

5. The UAV photovoltaic real-time inspection system based on the end-edge-cloud architecture according to claim 3 is characterized in that: The two-stage defect detection algorithm is specifically as follows: Taking infrared images as input, we first use a target recognition network to identify each photovoltaic module in the infrared image, and crop each sub-image containing the module based on the recognition results; then a classification network is used to identify the module defect types, which include normal modules, hot spot modules, severe hot spot modules, diode conduction modules, abnormal heating modules, distorted modules, and power-limited modules.

6. The UAV photovoltaic real-time inspection system based on the end-edge-cloud architecture according to claim 5 is characterized in that: The target recognition network adopts the YOLOv11s network, and the classification network adopts the ResNet34 network.

7. The UAV photovoltaic real-time inspection system based on the end-edge-cloud architecture according to claim 5 is characterized in that: The two-stage defect detection algorithm is trained in stages. In the first stage, a target recognition network is trained. With the infrared image as input and the annotated photovoltaic module block diagram as labels, the target recognition network is trained to identify the position of each photovoltaic module in the infrared image. In the second stage, with a separate photovoltaic module sub-image as input and the annotated photovoltaic module defect type as label, a classification network is trained to identify the defect type of the photovoltaic module sub-image.

8. The UAV photovoltaic real-time inspection system based on the end-edge-cloud architecture according to claim 1 is characterized in that: When the embedded platform uploads the image detection results to the cloud server, if the image size is larger than the threshold, the image is compressed.

9. The UAV photovoltaic real-time inspection system based on the end-edge-cloud architecture according to claim 1 is characterized in that: When the cloud server sends the results to the mobile phone, it constructs a JSON message containing the image address and annotation information; it uses MQTT to publish the JSON message to a specific topic; the mobile phone subscribes to the topic and receives the message, parses the received JSON message, extracts the image and annotation information from it, downloads the image file corresponding to the address through HTTP request, and displays the image and annotation data.

10. A control method for a UAV photovoltaic real-time inspection system based on an end-edge-cloud architecture according to claim 1, characterized in that: include: S1: Manually control the drone equipped with a thermal imaging camera to perform mapping tasks in the target work area, complete path planning and camera position setting, and all camera positions should cover the target work area; In S2, the drone performs daily inspection tasks, flying according to the planned track points, completing the task of collecting infrared and visible light images at the target position, and storing them on the gimbal camera end; In step S3, the embedded platform reads the data from the gimbal camera, first converts the visible light image into a grayscale image, and calculates the Laplacian variance of the grayscale image as a fuzzy score. If the fuzzy score is less than the fuzzy threshold, the corresponding infrared image is marked as fuzzy and uploaded directly to the cloud server without performing defect detection. Otherwise, the two-stage defect detection algorithm is called to identify the defect type of each photovoltaic module in the infrared image. The image detection results are then uploaded to the cloud server. S4, the cloud server pushes the results to the mobile phone through the message server, and generates a detection report after the flight mission is completed.

Images